DigitalRavagers/Experiment

product-launch-video - dispatch entry

Confirm the route before Step 0. This skill makes a video for a product being marketed / launched / promoted. If it's really a general (non-launch) site → video (site tour / showcase, not selling a product) → /website-to-video; a topic / concept with no product → /faceless-explainer; a GitHub PR → /pr-to-video; an existing video to caption / package → /embedded-captions · /graphic-overlays. Out of scope (decline, don't fake): live / at-render-time data (every value is baked in at author time), or footage / screenshots / an avatar that doesn't exist yet (HyperFrames can't record or capture). Routed here on a vague "make a video", or unsure product-vs-topic / launch-vs-general-site? Read /hyperframes first.

All artifacts are written to PROJECT_DIR = videos/<project-name>/ (created in Step 0). Paths below are relative to PROJECT_DIR. You (the orchestrator) run the Bash steps and dispatch the subagents; per-phase details live in the linked guides/agents/scripts — do not expand them here. Dispatch is harness-portable: before the first subagent dispatch, read <SKILL_DIR>/../hyperframes-core/references/subagent-dispatch.md once — it maps the dispatch verbs (parallel fan-out / background / wait) to your harness's primitives; a concurrency cap below N means waves of the cap size, never fewer scenes. This file is a binding runbook, not background reading: execute the steps in order and produce every phase artifact with its designated script or agent role — do not substitute a freestyle pipeline (hand-written narration, ad-hoc TTS calls, one hand-authored composition), and do not skip a pause step because the request seems clear. A step you cannot perform → stop and report; improvising past it breaks every downstream contract.

Phase	Execution	Primary artifact	Detailed flow
init	Bash directly	hyperframes.json	Step 0 (this file)
capture	Bash directly (hyperframes capture)	capture/extracted/tokens.json	phases/capture/guide.md
design-system	subagent	design-system/design.html + chunks/	agents/design-system.md
story-design	subagent	narrator_scripts.json	agents/story-design.md
audio	Bash directly (audio.mjs)	audio_meta.json	phases/audio/guide.md
visual-design	subagent	section_plan.md	agents/visual-design.md
prep	Bash directly (prep.mjs)	group_spec.json	scripts/prep.mjs header
captions	Bash directly (captions.mjs group -> html)	caption_groups.json + compositions/captions.html	scripts/captions.mjs header
scenes	N x subagent (parallel, one scene each)	compositions/scene_*.html	agents/hyperframes-scene.md
finalize	Bash prelude (wait-bgm + assemble + transitions + hoist + sfx-verify) -> finalize subagent	renders/video.mp4	Step 7 / agents/hyperframes-finalize.md

Prerequisites (install before first run)

macOS Apple Silicon or Linux x64:

brew install [email protected] node ffmpeg                   # On Linux, use the apt/dnf equivalent
npx hyperframes doctor                                  # One-time check that Chrome / dependencies are ready

• [email protected] — Homebrew Python, not system /usr/bin/python3 (PEP 668 blocks pip install otherwise); used by the MusicGen fallback
• node >= 18 + ffmpeg (audio.mjs uses ffprobe)
• Chrome downloads automatically on first npx hyperframes capture. hoist-videos.mjs (Step 7, runs only when a scene declares footage) reuses that cached Chrome; if it reports deps missing, run node <SKILL_DIR>/scripts/hoist-videos.mjs --ensure-deps once (~5s)

Optional API keys (unset -> local fallbacks; injection in Step 0.5; GEMINI_API_KEY ≡ GOOGLE_API_KEY):

Key	Used for	Default voice / fallback
HEYGEN_API_KEY (or hyperframes auth login)	TTS (cloud, with word-level timestamps)	voice: auto (first English starfish voice; override --voice)
ELEVENLABS_API_KEY	TTS (cloud; requires pip install elevenlabs)	voice 21m00Tcm4TlvDq8ikWAM (Rachel)
Neither, and not logged in	TTS	local Kokoro, voice am_michael (for non-English, pass --voice)
GEMINI_API_KEY (one key for both uses)	Capture vision caption + Lyria BGM	unset -> captions use DOM context only; BGM uses local MusicGen (first run downloads ~300 MB)

Flow

Step 0.0 - Confirm the brief (one round, then build)

Before Step 0, in one message confirm only what materially shapes the launch video and you can't infer — lead with a recommended default, skip anything the user already gave: the angle / focus (the product overall, a headline feature, an offer / CTA), length (default ~30-90s; up to ~3 min), and — if /hyperframes did not already set them — aspect (default 16:9; 9:16 for vertical / social) and language. The preset is derived from brand capture, not asked. For a fully specified request, skip this and build.

Step 0 - Initialize the video project

cwd is the agent workspace root; all video artifacts go in PROJECT_DIR = videos/<project-name>/.

Naming <project-name>: an explicit user-given directory wins; otherwise choose a short kebab-case name like <brand>-promo (never the workspace basename or a timestamp). From a URL, derive it from the domain/page title; the name is fixed once capture/ is written.

Initialization (only when $PROJECT_DIR/hyperframes.json does not exist):

PROJECT_DIR="${LAUNCH_VIDEO_DIR:-videos/<project-name>}"
mkdir -p "$(dirname "$PROJECT_DIR")"
npx hyperframes init "$PROJECT_DIR" --non-interactive --skip-skills --example=blank

hyperframes init drops a generic AGENTS.md / CLAUDE.md into $PROJECT_DIR; leave them in place but do not treat their generic guidance as run-time constraints — this skill is the source of truth.

Constraints (each violation breaks later phases):

• Do not run hyperframes init (or generate AGENTS.md / CLAUDE.md) in the workspace root; do not create a hyperframes/ subproject inside PROJECT_DIR.
• Every subagent dispatch context contains a PROJECT_DIR: <path> line; the subagent treats it as the project root.
• cwd discipline (master too): every Bash command runs as a (cd "$PROJECT_DIR" && ...) subshell — never bare cd "$PROJECT_DIR" && ... (persistent cwd drift makes later relative paths wrong).

Step 0.5 - API key guidance

Skip when $PROJECT_DIR/.env exists or context.log is non-empty. Otherwise detect what's configured (HeyGen TTS = $HEYGEN_API_KEY / $HYPERFRAMES_API_KEY / ~/.heygen/credentials; ElevenLabs / Gemini = their env keys), then always pause and ask — do not proceed on your own, even when a workable config is detected:

Detected: . Cloud keys are optional — without them, unconfigured providers fall back locally (TTS -> Kokoro unless HeyGen is configured; BGM -> MusicGen). Reply with:

• paste keys -> I will write them to $PROJECT_DIR/.env
• "go" -> proceed with what is configured now
• "skip" -> proceed with local fallbacks for anything unconfigured

Pasted keys -> Write/Edit $PROJECT_DIR/.env, one KEY=value per line (overwrite same-name keys, do not judge values). "go" / "skip" -> Step 1.

Step 1 - Capture (Phase 1)

1. Resolve SKILL_DIR and any explicit TARGET_URL from the prompt; ensure Step 0 ran.
2. Read $PROJECT_DIR/context.log if it exists and use the Resume table below to skip completed phases.
3. Classify the input (Step 1.0) to set CAPTURE and VO_MODE, then run the matching path. Both paths share the same downstream commands.

Step 1.0 - Classify the input (set CAPTURE + VO_MODE)

Input shape	What to do
Explicit URL in the prompt	TARGET_URL = that URL; CAPTURE=yes; no voice-over question (narration comes from the captured site). Path (A).
User pasted / pointed at a script or brief	(1) Save the verbatim text to $PROJECT_DIR/user_script.txt. (2) Ask the voice-over question once (below) → set VO_MODE. (3) Resolve a capture target from the script (below) → set CAPTURE.
A topic / brief with no script prose, no product	CAPTURE=no; no voice-over question. Path (B).

Voice-over question (only when the user supplied actual script prose) — ask one short line and wait:

Should I use your script verbatim as the voice-over, or restructure it into more screen-ready scene narration?

• Verbatim — keep the original wording; I only split scenes and pair visuals. Duration follows the script.
• Restructure — treat it as a brief and rewrite tighter narration, 1-2 sentences per scene.

VO_MODE = verbatim | restructure (default restructure); threaded to story-design in Step 2.

Resolve a capture target from the script — default to finding and crawling a site (real brand tokens beat preset fallbacks); skip only when the user opted out ("no web / text-only / no capture"). In order: (1) explicit http(s):// URL in the script → use it, announce, CAPTURE=yes; (2) clear brand/product name → WebSearch for the official site, confirm the resolved URL with the user in one line before crawling (decline / nothing credible → CAPTURE=no); (3) nothing derivable → CAPTURE=no.

Capture + user script coexist: the crawl supplies only brand tokens + assets + visual register; the narration spine stays user_script.txt (honored via VO_MODE), never the site's own copy.

(A) Capture path (CAPTURE=yes):

(cd "$PROJECT_DIR" && npx hyperframes capture "<TARGET_URL>" -o ./capture)

(B) No-capture path (CAPTURE=no) — synthesize a minimal capture package; downstream is identical. You (master) choose the preset (no site to infer from; pick from the 19 presets per user intent, or ask one short question). The full script/brief goes into visible-text.txt; colors:[] triggers the preset-palette fallback (fill colors only if the user named brand colors):

(cd "$PROJECT_DIR" && mkdir -p capture/extracted capture/assets)
(cd "$PROJECT_DIR" && cat > capture/extracted/tokens.json <<'JSON'
{ "title": "<brand/title>", "description": "<one-line>", "colors": [], "fonts": [], "headings": [], "sections": [], "ctas": [], "svgs": [], "cssVariables": {} }
JSON
)
(cd "$PROJECT_DIR" && echo '{}' > capture/extracted/design-styles.json)
(cd "$PROJECT_DIR" && printf '%s\n' "<full user script / brief>" > capture/extracted/visible-text.txt)

If the user already has a final narrator_scripts.json, place it in $PROJECT_DIR/; the Resume table skips story-design.

Shared downstream for both paths (Path B appends --style <chosen-preset> to build-design; Path A omits it for auto-inference):

(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/derive-context-pack.mjs --capture ./capture)
(cd "$PROJECT_DIR" && node <SKILL_DIR>/phases/design-system/scripts/build-design.mjs ./design-system --no-emit)   # Path B: append --style <chosen-preset>

Validation (stop and report if anything is missing; if capture/BLOCKED.md exists, the site blocked the crawl — follow the instructions inside it):

[ -s "$PROJECT_DIR/capture/extracted/tokens.json" ] && \
[ -s "$PROJECT_DIR/capture/extracted/design-styles.json" ] && \
[ -s "$PROJECT_DIR/capture/context_pack.md" ] && \
[ -s "$PROJECT_DIR/design-system/inference.json" ] && \
[ -d "$PROJECT_DIR/capture/assets" ] && echo ok || echo missing

Step 1b + Step 2 - design-system ∥ story-design (parallel fork)

Both subagents depend only on Step 1 artifacts and do not read each other's output — after capture validates, start them in parallel (two concurrent background dispatches, per the dispatch adapter); do not serialize:

• design-system: prompt = full agents/design-system.md + ## Dispatch context with SKILL_DIR / PROJECT_DIR / Target URL + the full text of design-system/inference.json inlined via cat (~2-4 KB, saves the subagent one Read).

• story-design: prompt = full agents/story-design.md + ## Dispatch context:

  SKILL_DIR: <absolute path>
  PROJECT_DIR: <video project root>
  Schema validator: <SKILL_DIR>/scripts/validate-narrator.mjs
  Design DNA: ./design-system/inference.json   # read site_dna once to set the narrative register
  Provided script: ./user_script.txt   # ONLY when the user supplied a script; omit the line otherwise
  Voice-over mode: <verbatim | restructure>   # pair with Provided script; omit otherwise
  Script style: Keep each scene's script concise - 1-2 sentences, no more than 20 words   # suspended in verbatim mode (length follows the script)
  Orientation: <landscape | portrait | square>   # from the user's aspect (16:9→landscape, 9:16→portrait, 1:1→square; default landscape). Echoed verbatim into narrator_scripts.orientation → sets the canvas for the whole pipeline
  

Step 3 - Audio (Phase 2.5)

After story-design returns (narrator_scripts.json exists) — audio does not wait for design-system:

(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/audio.mjs \
  --narrator-scripts ./narrator_scripts.json \
  --hyperframes . \
  --out ./audio_meta.json \
  --lyria-recipe <SKILL_DIR>/phases/audio/lyria-recipe.py)

BGM runs detached in the background when available ($GOOGLE_API_KEY → Lyria cloud; else installed transformers torch soundfile numpy → local MusicGen, first run ~300 MB) and is silently skipped otherwise; all flags (--voice / --provider / --no-bgm / ...) are documented at the top of audio.mjs.

• exit 0 -> voice + transcribe complete (BGM may still be running; audio_meta.json records bgm_log / bgm_pid), continue.
• exit 1 -> zero scenes produced voice; report and stop.

Step 4 - Visual design (Phase 3)

Join point: design-system/chunks/index.json + narrator_scripts.json + audio_meta.json all exist. Build one dispatch packet (the subagent reads it once, zero extra Reads):

# Dispatch packets live in $PROJECT_DIR/.dispatch/ (transient; safe to delete after the run).
# NEVER use a fixed /tmp path: it persists across runs/projects, so a failed write silently
# reuses another project's stale packet and contaminates every worker.
mkdir -p "$PROJECT_DIR/.dispatch"
DP="$PROJECT_DIR/.dispatch/vd-dispatch.txt"
{
  # Section order is deliberate: contracts first, static references middle, work items last
  echo "## Design chunks"
  (cd "$PROJECT_DIR" && cat design-system/chunks/index.json \
    design-system/chunks/composition-hints.md design-system/chunks/voice.md \
    design-system/chunks/tokens.css design-system/chunks/easings.js 2>/dev/null)
  echo "## Effects catalog";  cat <SKILL_DIR>/phases/visual-design/effects-catalog.md
  echo "## Blueprints index"; cat <SKILL_DIR>/phases/visual-design/blueprints-index.md
  echo "## Design rules";     cat <SKILL_DIR>/phases/visual-design/rules/{typography,color-system,composition,motion-language}.md
  echo "## SFX library";      cat <SKILL_DIR>/assets/sfx/manifest.json
  echo "## Narrator scripts"; (cd "$PROJECT_DIR" && cat narrator_scripts.json)
  echo "## Audio meta";       (cd "$PROJECT_DIR" && cat audio_meta.json 2>/dev/null)   # optional; overrides Duration on >10% drift
} > "$DP"
# Guard: a partially-failed build must fail LOUDLY here, not downstream in the subagent
grep -q '^## Narrator scripts' "$DP" || { echo "FATAL: vd-dispatch.txt incomplete — rebuild before dispatching"; }

# Captions planning hint for the Captions: dispatch line below
(cd "$PROJECT_DIR" && node -e 'try{const m=require("./audio_meta.json");process.stdout.write(Object.values(m.scenes||{}).some(s=>s.wordsPath)?"enabled":"disabled")}catch{process.stdout.write("enabled")}')

Dispatch the subagent: prompt = full agents/visual-design.md + ## Dispatch context (copy verbatim, do not digest):

SKILL_DIR: <absolute path>
PROJECT_DIR: <video project root>
Schema validator: <SKILL_DIR>/scripts/validate-section.mjs
Canvas: <width>×<height>   # 1920×1080 default; 1080×1920 portrait / 1080×1080 square when narrator_scripts.orientation says so
Captions: <enabled | disabled>   # the node -e hint above; enabled => plan keeps key content in the upper ~83%
Dispatch packet: <PROJECT_DIR>/.dispatch/vd-dispatch.txt

The Captions: line is an optimistic hint; the authoritative gate is group_spec.captions_enabled from Step 5 prep (mismatch is safe — Step 6/7 keep-out always follows group_spec).

Step 5 - prep (deterministic, NO subagent)

After section_plan.md exists:

(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/prep.mjs \
  --section-plan ./section_plan.md \
  --narrator-scripts ./narrator_scripts.json \
  $( [ -f audio_meta.json ] && echo "--audio-meta ./audio_meta.json" ) \
  --rules-dir <SKILL_DIR>/../hyperframes-animation/rules \
  --capture ./capture \
  --design-system ./design-system \
  --hyperframes . \
  --sfx-lib <SKILL_DIR>/assets/sfx \
  --out ./group_spec.json)

• exit 0 -> append the stdout summary to $PROJECT_DIR/context.log.
• exit 1 -> stderr names the failing scene + anchor; re-dispatch visual-design (Step 4) with the error passed through.

Step 5.5 + Step 6 - Captions (deterministic) + scene worker fan-out

Captions are two Bash scripts, no subagent (run after prep, before fan-out):

(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/captions.mjs group \
  --group-spec ./group_spec.json --hyperframes . \
  --tokens design-system/chunks/tokens.css --out ./caption_groups.json)

(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/captions.mjs html \
  --hyperframes . --groups ./caption_groups.json \
  --tokens design-system/chunks/tokens.css \
  --inference design-system/inference.json \
  --out compositions/captions.html)

exit 0 = normal. captions: skipped (<reason>) = legal skip — no captions.html, assemble will not mount track 12; continue. Do not run npx hyperframes lint <file> on captions.html (lint takes a project directory; a file path exits 1).

Scene worker fan-out: read group_spec.json.groups[] for worker count N and group_spec.captions_enabled for the Captions: flag, then build the per-worker dispatch packets and start N workers in parallel (concurrent background dispatches; a harness concurrency cap below N means waves of the cap size until all N scenes exist — one scene per worker either way, never fewer scenes):

# Same rule as Step 4: packets go in $PROJECT_DIR/.dispatch/, never a fixed /tmp path
# (a stale /tmp file from a previous project survives a failed write and silently
# poisons every worker with the wrong design system).
mkdir -p "$PROJECT_DIR/.dispatch/scene-dispatch"
# Shared header (identical for every worker), computed once:
# `## Film direction` = the film-level invariants from group_spec.film_direction
# (palette system / motion defaults + budget / ambient system / negative list);
# each scene's creative_brief carries only scene-specific deltas on top of it.
{
  echo "## Film direction"
  (cd "$PROJECT_DIR" && node -p 'JSON.parse(require("fs").readFileSync("group_spec.json","utf8")).film_direction || ""')
  echo "## Tokens / easings / voice"
  (cd "$PROJECT_DIR" && cat design-system/chunks/tokens.css design-system/chunks/easings.js design-system/chunks/voice.md 2>/dev/null)
} > "$PROJECT_DIR/.dispatch/scene-shared.txt"
# Guard BEFORE fan-out: header structure + the project's own brand token must both be present;
# a contaminated packet here costs a full re-author round across every affected worker.
grep -q '^## Film direction' "$PROJECT_DIR/.dispatch/scene-shared.txt" && \
  grep -q -- '--brand-primary' "$PROJECT_DIR/.dispatch/scene-shared.txt" || \
  { echo "FATAL: scene-shared.txt incomplete/stale — rebuild before dispatching workers"; }
# Per-worker packet: shared header + that worker's Scenes YAML -> $PROJECT_DIR/.dispatch/scene-dispatch/w<N>.txt

Each worker's prompt = full agents/hyperframes-scene.md + ## Dispatch context with: SKILL_DIR / PROJECT_DIR / Worker ID / Composition width + Composition height (= group_spec.width/height) / Captions: <enabled|disabled> / Dispatch packet: <PROJECT_DIR>/.dispatch/scene-dispatch/w<N>.txt, plus the shared header body (## Film direction + ## Tokens / easings / voice) and the worker's Scenes: list copied verbatim from group_spec.json.groups[i].scenes[<sid>] (scene_id / effects / rule_paths / assetCandidates / estimatedDuration_s / voicePath / blueprint / design_chunks / creative_brief). When Captions: enabled, also pass Caption band top y = height − round(height × 0.1667) and Foreground max y = Caption band top y − 20 (landscape → 900 / 880; portrait → 1600 / 1580). design_chunks: null (anomaly already reported by prep) -> the worker falls back to reading design-system/design.html.

After all workers return, run the static composition gate (scans compositions/scene_*.html per group_spec.scene_ids; captions.html is covered by its own self-lint + Step 7 whole-project lint):

(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/check-compositions.mjs \
  --hyperframes . \
  --group-spec ./group_spec.json)

• 0 -> continue to Step 7.
• 1 -> stderr names the violating scene + rule; re-dispatch that worker (do not Edit in the master).

Step 7 - Assembly prelude + finalize (Phase 4c)

(1) Deterministic Bash prelude (each script documents its internals in its own header; you only branch on exit codes):

(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/wait-bgm.mjs \
  --audio-meta ./audio_meta.json \
  --hyperframes . \
  --timeout-ms 120000 \
  --interval-ms 2000)
(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/assemble-index.mjs --group-spec ./group_spec.json --hyperframes .)
(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/transitions.mjs inject --group-spec ./group_spec.json --hyperframes .)
(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/transitions.mjs verify --group-spec ./group_spec.json --index ./index.html)
(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/hoist-videos.mjs --group-spec ./group_spec.json --hyperframes .)
(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/verify-output.mjs sfx --group-spec ./group_spec.json --index ./index.html)

No agent hand-writes index.html or manually checks BGM. Exit-code branches:

• assemble exit 1 -> names a scene (root data-duration ≠ group_spec, or file missing) = worker contract break -> re-dispatch that worker (Step 6), then rerun this step.
• transitions inject/verify exit 1 -> injector bug (prep already validated transitions[]) -> report for investigation; do not roll back workers.
• hoist-videos exit 1 -> an invalid data-video-src declaration (stderr names scene + reason); Edit the scene file (or re-dispatch for a real relayout), rerun this step. exit 2 -> run node <SKILL_DIR>/scripts/hoist-videos.mjs --ensure-deps from the workspace root, rerun.
• sfx-verify exit 1 -> assembler bug -> report for investigation.

(2) Dispatch the finalize subagent: prompt = full agents/hyperframes-finalize.md + ## Dispatch context:

SKILL_DIR: <absolute path>
PROJECT_DIR: <video project root>
Render quality: high     # or draft / standard
Captions: <enabled | disabled>   # = group_spec.captions_enabled
BGM: <one-line wait-bgm verdict, e.g. "ready (lyria)" / "skipped (no key)" / "timeout">
Film direction: |        # = group_spec.film_direction (film-level invariants the briefs assume)
  <verbatim>
Scenes:                  # one line per scene, copied verbatim from group_spec.json
  - { scene_id, start_s, estimatedDuration_s, effects: [...], creative_brief: |
      <Phase 3 prose for this scene> }

Finalize runs lint+validate, takes one contact-sheet look, fixes what the pixels show, renders, and verifies — its prompt owns that flow. Outcomes:

• Reports the verified mp4 + fixes in place -> complete.
• STOP (a scene needs real recomposition — exception, not default) -> re-dispatch that worker (Step 6) with normal dispatch context + a ## Repair context block carrying finalize's verbatim findings and the Captions: flag -> rerun (1) -> re-dispatch finalize. Same finding survives two rounds -> stop and surface to the user.

Completion report

Summarize per phase from context.log + each step's stdout: capture URL / asset counts, preset, archetype, scene count + total duration, transitions, gate status, fixes in place, final mp4 path + bytes + duration.

Offer a live preview — never auto-open one. The deliverable is the mp4. Do NOT run hyperframes preview / play during any earlier phase (a mid-run preview shows half-edited compositions and dies with that phase's server). Only when the user asks, after the render:

(cd "$PROJECT_DIR" && npx hyperframes preview)   # Studio UI, e.g. http://localhost:3002/#project/<project-name>
(cd "$PROJECT_DIR" && npx hyperframes play)       # or a plain shareable player at http://localhost:<port>

Report the actual URL with the real port. Flags live in the hyperframes-cli skill.

---

Resume table

Read $PROJECT_DIR/context.log and resume from the first missing artifact:

State	Continue from
log missing or empty	Full pipeline
capture/extracted/tokens.json missing	Rerun Step 1 (capture + derive-context-pack + build-design.mjs --no-emit)
tokens.json exists, design-system/inference.json missing	Rerun only build-design.mjs --no-emit (deterministic, seconds)
inference.json exists, but design.html or narrator_scripts.json missing	Step 1b/2 fill-in: dispatch whichever subagent's artifact is missing (both missing -> both in parallel)
narrator_scripts.json exists, audio_meta.json missing	Step 3 (audio)
audio_meta.json exists, section_plan.md missing	Step 4 (visual-design)
section_plan.md exists, group_spec.json missing	Step 5 (prep)
group_spec.json exists, compositions/scene_*.html missing or caption_groups.json missing	Step 5.5+6: run the captions scripts first, then dispatch workers for whichever scenes are missing, in parallel. Captions-ran criterion = caption_groups.json exists (a legal skip writes no captions.html; keying on captions.html would re-skip forever)
All compositions/scene_*.html exist + captions state decided, renders/video.mp4 missing	Step 7: rerun the full Bash prelude (overwrite index.html — upstream scenes may have changed), then dispatch finalize
renders/video.mp4 exists	Report completed and stop